fix: #8612
For checkpoint experiment in some case users might want to give it a early
stopping to reduce variance. But currently if we interrupt/kill the
experiment it will be marked as failed, and all of the completed
checkpoints will be removed as we cleanup all the running directly after
the process failed.
1. We raise CheckpointKilledError intead of StageCmdFailed error if at
least one checkpoint had been commited.
2. Temp_dir executor will continue collecting data if the Checkpoint
stage was interrupted.
3. Raise warnings if a checkpoint stage was incomplete and the other
stages were not forwarded.
4. Add a new functional test for this
This is a very early version of test utils for dvc plugins. This module
consists of existing tests with minimal changes and is going to evolve
a lot as our plugins are getting filled in and pytest-test-utils matures.
Intermediate PR in removing path_info from our code base.
It is completely removed from the main codebase and is still
being used in tests, which will be migrated in the followups.
For fs-specific simple path manipulation, we've introduced
`fs.path` helper (similar to `os.path`).
Fixes#6774
* Push running checkpoint to remote
Fix#6182
* Update dvc/repo/experiments/executor/base.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* Get the branch name
* Finish this PR
1. move env name to dvc/env.py.
2. add some tests for it.
* Update after finally commit
* Some problems found in the review
1. change the behavior of self remote
2. do not use string DVC_EXP_AUTO_PUSH
3. downgrade the logger level in auto push
4. use full branch ref
* Update dvc/repo/experiments/executor/base.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* add a remote check before auto push
* Split current one env into two
* Rename DVC_EXP_CHECKPOINT_PUSH to DVC_EXP_AUTO_PUSH
* value error
* Remove getenv_bool switch to env2bool
* Better on handling Dulwich exceptions
* Name changing removing prints
* Update dvc/env.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* Some changes in review
* Add validation to repo URL
* Add new type of SCMError
* Update dvc/repo/experiments/executor/base.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* Use spy tests call counts and args
* Update dvc/scm/git/backend/base.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* Update dvc/scm/git/backend/dulwich.py
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove validate_git_repo to validate_git_remote
Co-authored-by: Peter Rowlands (변기호) <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* apply: allow applying exp with multiple possible branches
* default to new checkpoint branch after applying ambiguous commit
* output shas in addition to exp names by default
* update show tests
* git: allow force checkout in detach_head context
* exp run: error when trying to resume from a non-head checkpoint without
modifications
* add test for apply/resume from intermediate checkpoint
* list some possible heads when multiple matches
* experiments: move executor into a module
* experiments: separate executor into base/local
* rename LocalExecutor to TempDirExecutor, run in .dvc/tmp by default
* git: restore HEAD using reset not checkout in detached_head()
* experiments: support calling BaseExecutor.reproduce directly for
workspace runs
* experiments: run single experiments in local workspace by default
* update experiments tests for workspace changes
* experiments: update checkpoints for workspace runs
* update checkpoints tests
* update test_show
* add force param for apply
* exp apply: add --force option
* exp run/res: add --temp option to force tempdir execution for single
experiments
* force apply in non-apply specific tests
* experiments: commit to refs/exps/... instead of branch heads
* update exp show
* restore checkpoint_resume kwarg
* refactor new() to use experiments stash
* experiments: use refspecs to populate and collect executors
* experiments: use new stash and merge on exp checkout
* experiments: fix executor chdir
* experiments: update run and show
* skip checkout after reproduce
* checkpoints: use experiments refspecs
* check if workspace was stashed before applying
* exp show: unify checkpoint/standalone experiment collection
* git: move iter_refs to backend
* update for scm changes
* experiments: replace exp checkout in favor of exp apply
* git: handle rev_parse errors in resolve_commit
* update refs containing and baseline check
* fix sigint handling for checkpoints
* update checkpoint exists behavior
* fix checkpoint baseline behavior
* update tests
* experiments: update gc
* update show tests
* fix type checking errors
* update diff
* scm: add diff, describe
* experiments: generate human readable names
* exp run/resume: allow user specified exp name with -n/--name
* update tests
* experiments: add `exp branch` for promoting experiment to full branch
* fix test
* fix missing file
* experiments: explicitly close scm instances before cleaning up tmp dir
* experiments: Allow branching from checkpoints with modified hyper params
* experiments: initial checkpoints tree functionality for `exp show`
* go back to flat/sortable view
* add test for resume w/modified params
* tests: move experiments fixtures into conftest
* add show checkpoints test
* test checkpoint table rendering
* remove dead code
* test show for checkpoint branches
* experiments: refactor regular/checkpoint experiment separation
* fix scm_checkout pull bug
* experiments: return stages from executor.reproduce
* since executor has to move to threadpool execution anyways, we no
longer have to work around the pickle repro result limitation
* refactor new/resume behavior
* refactor checkpoints tests
* use -f to force checkpoints reset/restart
* checkpoints: use 'exp run' and 'exp res[ume]' commands
* stage: all outs can be missing for checkpoint stages
* regular outputs may not be written at the time a checkpoint is
generated, but subsequent checkouts of that stage should not fail due
to the missing outs
* commit stage on checkpoint generation otherwise non-persist cached outs
are lost
* checkpoints: wait for stage proc to end after killing via terminate()
* stage: fix issues with killing checkpoint stage
* refactor signaling (use threading.Event instead of Condition)
* fix windows bug (kill all child processes via taskkill)
* experiments: refactor executor collection
* tests: move checkpoint tests into separate module
* refactor experiment tests to use stage fixture
Peter Rowlands
skshetry
dberenbaum
karajan1001
Ruslan Kuprieiev
Saugat Pachhai (सौगात)
Batuhan Taskaya
Sami Jawhar